md/raid10: locking changes for 'enough()'.
authorNeilBrown <neilb@suse.de>
Tue, 11 Jun 2013 04:57:09 +0000 (14:57 +1000)
committerNeilBrown <neilb@suse.de>
Thu, 13 Jun 2013 22:10:27 +0000 (08:10 +1000)
As 'enough' accesses conf->prev and conf->geo, which can change
spontanously, it should guard against changes.
This can be done with device_lock as start_reshape holds device_lock
while updating 'geo' and end_reshape holds it while updating 'prev'.

So 'error' needs to hold 'device_lock'.

On the other hand, raid10_end_read_request knows which of the two it
really wants to access, and as it is an active request on that one,
the value cannot change underneath it.

So change _enough to take flag rather than a pointer, pass the
appropriate flag from raid10_end_read_request(), and remove the locking.

All other calls to 'enough' are made with reconfig_mutex held, so
neither 'prev' nor 'geo' can change.

Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid10.c

index 3c6b193..5169ed2 100644 (file)
@@ -97,7 +97,7 @@ static int max_queued_requests = 1024;
 
 static void allow_barrier(struct r10conf *conf);
 static void lower_barrier(struct r10conf *conf);
-static int enough(struct r10conf *conf, int ignore);
+static int _enough(struct r10conf *conf, int previous, int ignore);
 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
                                int *skipped);
 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
@@ -392,11 +392,9 @@ static void raid10_end_read_request(struct bio *bio, int error)
                 * than fail the last device.  Here we redefine
                 * "uptodate" to mean "Don't want to retry"
                 */
-               unsigned long flags;
-               spin_lock_irqsave(&conf->device_lock, flags);
-               if (!enough(conf, rdev->raid_disk))
+               if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
+                            rdev->raid_disk))
                        uptodate = 1;
-               spin_unlock_irqrestore(&conf->device_lock, flags);
        }
        if (uptodate) {
                raid_end_bio_io(r10_bio);
@@ -1632,9 +1630,17 @@ static void status(struct seq_file *seq, struct mddev *mddev)
  * Don't consider the device numbered 'ignore'
  * as we might be about to remove it.
  */
-static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
+static int _enough(struct r10conf *conf, int previous, int ignore)
 {
        int first = 0;
+       int disks, ncopies;
+       if (previous) {
+               disks = conf->prev.raid_disks;
+               ncopies = conf->prev.near_copies;
+       } else {
+               disks = conf->geo.raid_disks;
+               ncopies = conf->geo.near_copies;
+       }
 
        do {
                int n = conf->copies;
@@ -1644,25 +1650,31 @@ static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
                        if (conf->mirrors[this].rdev &&
                            this != ignore)
                                cnt++;
-                       this = (this+1) % geo->raid_disks;
+                       this = (this+1) % disks;
                }
                if (cnt == 0)
                        return 0;
-               first = (first + geo->near_copies) % geo->raid_disks;
+               first = (first + ncopies) % disks;
        } while (first != 0);
        return 1;
 }
 
 static int enough(struct r10conf *conf, int ignore)
 {
-       return _enough(conf, &conf->geo, ignore) &&
-               _enough(conf, &conf->prev, ignore);
+       /* when calling 'enough', both 'prev' and 'geo' must
+        * be stable.
+        * This is ensured if ->reconfig_mutex or ->device_lock
+        * is held.
+        */
+       return _enough(conf, 0, ignore) &&
+               _enough(conf, 1, ignore);
 }
 
 static void error(struct mddev *mddev, struct md_rdev *rdev)
 {
        char b[BDEVNAME_SIZE];
        struct r10conf *conf = mddev->private;
+       unsigned long flags;
 
        /*
         * If it is not operational, then we have already marked it as dead
@@ -1670,18 +1682,18 @@ static void error(struct mddev *mddev, struct md_rdev *rdev)
         * next level up know.
         * else mark the drive as failed
         */
+       spin_lock_irqsave(&conf->device_lock, flags);
        if (test_bit(In_sync, &rdev->flags)
-           && !enough(conf, rdev->raid_disk))
+           && !enough(conf, rdev->raid_disk)) {
                /*
                 * Don't fail the drive, just return an IO error.
                 */
+               spin_unlock_irqrestore(&conf->device_lock, flags);
                return;
+       }
        if (test_and_clear_bit(In_sync, &rdev->flags)) {
-               unsigned long flags;
-               spin_lock_irqsave(&conf->device_lock, flags);
                mddev->degraded++;
-               spin_unlock_irqrestore(&conf->device_lock, flags);
-               /*
+                       /*
                 * if recovery is running, make sure it aborts.
                 */
                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
@@ -1689,6 +1701,7 @@ static void error(struct mddev *mddev, struct md_rdev *rdev)
        set_bit(Blocked, &rdev->flags);
        set_bit(Faulty, &rdev->flags);
        set_bit(MD_CHANGE_DEVS, &mddev->flags);
+       spin_unlock_irqrestore(&conf->device_lock, flags);
        printk(KERN_ALERT
               "md/raid10:%s: Disk failure on %s, disabling device.\n"
               "md/raid10:%s: Operation continuing on %d devices.\n",
@@ -1791,7 +1804,7 @@ static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
                 * very different from resync
                 */
                return -EBUSY;
-       if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
+       if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
                return -EINVAL;
 
        if (rdev->raid_disk >= 0)